Monitoring a person's dietary behavior has many health applications (e.g. weight management). Existing approaches for automatic eating detection using wearable sensors often require custom hardware that may not be practical as their usability and energy efficiency have not been validated. In this paper, we propose to use off-the-shelf Bluetooth headsets to unobtrusively monitor and detect users' eating episodes by analyzing the chewing sound. The challenges of using commodity acoustic hardware include the limited sampling rate that reduces feature fidelity and the impact of environment noise in real-world settings. Our experimental results show that the traditional kernel-based approach using Support Vector Machine (SVM) could achieve 94-95% classification accuracy in lab settings, though the detection performance quickly degraded to 65-76% for in-the-field testing. We then propose a novel Deep Learning based classification technique that drastically improved detection accuracy to 77-94% despite the existence of ambient noise. To the best of our knowledge, this is the first study to use Bluetooth headsets to detect eating episodes and to use Deep Learning to significantly increase classification performance in this context. We believe that the adoption of readily available and low-cost consumer devices provides a foundation for practical deployment of automated dietary monitoring applications.